1 / 10

Feedback Request Strawman: Using .1Qau for Load Monitoring in DC Networks

Feedback Request Strawman: Using .1Qau for Load Monitoring in DC Networks. Mitch Gusat IBM ZRL, Aug. 28, 2008. What is Fb_Rq? On demand status info. Problem: Monitoring, app-level (L4+) performance profiling, runtime load balancing, adaptive routing...

oona
Download Presentation

Feedback Request Strawman: Using .1Qau for Load Monitoring in DC Networks

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Feedback Request Strawman:Using .1Qau for Load Monitoring in DC Networks Mitch Gusat IBM ZRL, Aug. 28, 2008

  2. What is Fb_Rq? On demand status info Problem: Monitoring, app-level (L4+) performance profiling, runtime load balancing, adaptive routing... Solution: Build on the investment in .1Qau-compliant switches => Deliver the full/available feedback to sources ! (before congestion arises) Benefits • Speed: L2 feedback • Accuracy: Q info is already known to CP for QCN Fb. Ship it to RP! • Communicate Fb up the L3-7 stack. Use Flow-/RP-ID (?).

  3. Fb_Rq Basics • Monitoring options • Proactive: RP-initiated => RP autonomosly issues Fb_Rq • [Reactive: CP-initiated => RP begins to ping after QCN CNM] • Single CP (reflect) vs. path (reflect reply & fwd request ) Fb • Stateless (anon.) vs. statefull CP (pings counted per FlowID)

  4. Fb_Rq Strawman’s Steps • RP: injects Fb_Rq pkt w/ L2 flag and Seq./Flow/RP-ID • CP: receives Fb_Rq • sets Psample=1 (or disregards if busy or in “silent” mode) • dumps queue status info (see next) • sends Fb_Rp (CNM-like) back to originating SRC • optionally also forwards Fb_Rq if DST != local CP => path profiling (multi-pathing issues)

  5. Extended Queue Status (EQS) • Prio, Qsize, Qeq, Qoff, Qdelta + options • PingCnt: # of pings (from any FlowID) RX-ed since the last change of q’ sign • marks one monotonic episode (of Q qrowth or drain) • If aggregate per CP, it can provide HSD/no-sharers (if RP maintains its own PingCnt) • TXCnt: # of pkts forwarded since the last change of q’ sign  as proxy hint for avg. service rate • ... • [additional info, e.g. pointer to a complete CP “brain dump”]

  6. Reactive Fb_Rq Operation (animated) • Qeq exceeded • Send QCN Fb to RP • Reactive probing is triggered by QCN frames • hence only rate-limited flows are probed • Insert one Fb_Rq ping every n KB of data sent per flow, e.g. n = 750 KB • Single CP probing: CPID of probes = destination MAC • Pro-active probing needs no CNM, but n should be based on actual load and delay Fb_Rq • Fb_Rq arrives @ CP • Insert EQS data • Return Fb_Rq to source Switch 1 Switch 2 src CNM • Initial CNM arrives @ RP • Install RL • Inject Fb_RQ Switch 3 dst • Fb_Rq reply arrives @ RP • EQS feedback is extracted and sent up to L3 IBM Research GmbH, Zurich

  7. Conclusion Q: What is being enabled? A: Anticipate overload “see it coming” • Potential for *early* custom response to congestion thru application specific logic: • e.g. app-driven adaptive routing, • task migration, • collectives (MPI mcast, combining ops, locks), • LB-ing engines, • scheduling hints: "optimize for latency" or "optimize for throughput“, • control of new session admittance (postpone a bkup after the trading rush), • redundant data placement, • ... IBM Research GmbH, Zurich

  8. BKUP

  9. Why bother about Q’? • Delay: queuing delay-dominated RTT destabilizes CM in large DC’s • additionally the RP delay further reduces RTT budget (see 21st Aug. call) • Oscillations • with quick On/Off congestion episodes false recoveries are frequent (presented in 2007) • when RTT > 0.5-1ms Qoff is (much) less significant than Qdelta • Q’ provides additional info • Luckily Q’ (aka Qdelta) is available @ CP • Q’ potential usage • Q’ marks monotonical periods: queue backlogging / draining • can provide HSD (N sharers) Fb • extends the state space {q,q’} for tighter RL control @ RP • enables RTT compensation IBM Research GmbH, Zurich

  10. Pkt. Format • TBD... • Enhanced CNM: add EQS to CNM’s Fb (QCN)

More Related